Safe Navigation in Uncertain Crowded Environments Using Risk Adaptive CVaR Barrier Functions

Xinyi Wang1, Taekyung Kim1, Bardh Hoxha2, Georgios Fainekos2, Dimitra Panagou1,
1Department of Robotics, University of Michigan, 2Toyota North America Research & Development
Paper Video Code

Abstract

We propose a risk-adaptive approach based on the Conditional Value-at-Risk Barrier Function (CVaR-BF), where the risk level is automatically adjusted to accept the minimum necessary risk, ensuring CVaR safety is guaranteed at least a pre-defined threshold while improving optimization feasibility under uncertainty. Additionally, we introduce a dynamic zone-based barrier function which characterizes the collision likelihood by evaluating the relative state between the robot and the obstacle. It expands the available adjustment space for the risk level while maintaining the desired probabilistic safety guarantee. By integrating risk adaptation with this new function, our approach enables the robot to proactively avoid obstacles in highly dynamic environments.

Motivation

Dolly zoom effect
Matting example

(a) A fixed risk level is not flexible enough: A low risk tolerance enhances safety but can render the optimization infeasible, whereas a high risk tolerance improves feasibility at the expense of safety.
(b) In highly dynamic scenarios, where obstacles move unpredictably and rapidly, the robot requires sufficient time and space to respond and adjust its risk level.
(c) Our method dynamically adjusts the risk level within an extended risk range to maintain feasibility while ensuring user-defined probabilistic safety. The robot proactively modifies its trajectory before approaching obstacles, but only when necessary, thus avoiding unnecessary conservatism.

Method Overview

Method Overview Figure
Teaser GIF

Animations for different number of obstacles

Comparison results on 20 obstacles with noise σ=0.025

Real-World Dataset Validation

We evaluated the proposed CVaR-BF optimization as a safety filter for RL-based methods, such as CrowdNav++ (serving as the nominal controller), on a real-world pedestrian trajectory dataset (Univ Scenario)  [1]. Experimental results demonstrate that the RL agent alone frequently failed to avoid collisions in out-of-distribution (OOD) scenarios, where the robot radius was larger than during training. In contrast, with our safety filter, the RL agent navigated safely even in these challenging OOD settings.
#Obs. = 36   |   rrobot = 0.6m, robs = 0.2m   |   vmax = 1 m/s, vmaxobs = 1.7 m/s
[1] Biswas, S. et al., SocNavBench: A Grounded Simulation Testing Framework for Evaluating Social Navigation.
CrowdNav++
CrowdNav++
CrowdNav++ & CVaR-BF (Distance)
CrowdNav++ & Proposed

Video

BibTeX

@article{wang2025safe,
  title={Safe Navigation in Uncertain Crowded Environments Using Risk Adaptive CVaR Barrier Functions},
  author={Wang, Xinyi and Kim, Taekyung and Hoxha, Bardh and Fainekos, Georgios and Panagou, Dimitra},
  conference={IROS},
  year={2025}
}